Enhanced regulatory sequence prediction using gapped k-mer features.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• K-mers are widely used for DNA and protein sequence analysis.
• Increasing k-mer length for longer features leads to sparse data and noisy frequencies, limiting statistical learning.
• Existing methods struggle with large-scale genomic applications due to k-mer limitations.

Purpose of the Study:

• To develop a robust method for sequence analysis that overcomes k-mer limitations.
• To introduce gapped k-mers (gkm) and a new classifier, gkm-SVM, for improved feature representation.
• To enhance the accuracy of predicting functional genomic regulatory elements and enhancers.

Main Methods:

• Introduced gapped k-mers (gkm) as an alternative feature set.
• Developed the gkm-SVM classifier for sequence classification.
• Implemented an efficient tree data structure for kernel matrix computation for large-scale applications.
• Utilized robust estimation for k-mer frequencies.

Main Results:

• Gkm-SVM significantly improves the accuracy of predicting functional genomic regulatory elements and tissue-specific enhancers, with precision increased by up to a factor of two.
• Gkm-SVM consistently outperforms the original k-mer-SVM on human ENCODE ChIP-seq datasets.
• Demonstrated the general utility of the gkm-SVM method using a Naïve-Bayes classifier.

Conclusions:

• Gapped k-mers and gkm-SVM offer a more robust and accurate approach to sequence analysis compared to traditional k-mers.
• The developed methods are efficient and scalable for large-scale genome-wide applications.
• The gkm-SVM approach is broadly applicable to various sequence classification problems beyond regulatory element analysis.